Screening for conjunctions

What you will accomplish

A conjunction screen you can act on, and a clear view of which parts of the answer are geometry and which are assumptions you supplied.

Two things in this workflow produce alarming numbers that mean nothing. Both are shown here with real output.

Prerequisites

• A constellation.
• Optionally a TLE catalog to screen against.

Steps

orbitforge conjunction screen \
  --constellation ph1.json \
  --duration-hours 24 --threshold-km 10 --step-seconds 60

Screened 60 objects (all pairs) over 24.0 h: 0 events under 10 km.

orbitforge conjunction screen \
  --constellation ph1.json \
  --catalog stations.tle \
  --duration-hours 24 --threshold-km 10 --step-seconds 60

A well-formed Walker does not conflict with itself

The intra-constellation screen returns nothing, and raising the threshold does not change that:

Note also that conjunction screen has no --model flag. It does not accept a propagation model the way simulate and coverage do, so plan the epoch discipline for a TLE-derived fleet accordingly.

1042 events, none of them collisions

Screening a TLE-imported fleet against the catalog it came from:

Screened 24 objects vs 24 catalog entries over 24.0 h: 1042 events under 10 km.
  iss-49271 / FREGAT DEB: TCA 2026-06-26T02:10:14Z miss 1.164 km at 0.00 km/s (RIC [-0.000, -1.164, -0.000])
  iss-53239 / CSS (WENTIAN): TCA ... miss 1.374 km at 0.00 km/s
  iss-53239 / TIANZHOU-10: TCA ... miss 1.375 km at 0.00 km/s

Three signatures identify them, and all three are in the output:

1. Relative speed near zero. The decisive one.
2. Purely along-track separation. The cross-track component of the RIC offset comes out around a nanometer, which is numerically zero.
3. The same pair repeating at the orbital period. The 1042 events come from only 64 distinct pairs, and the most frequent produced 17 events each over 24 hours, which is one per revolution.

The screen did exactly what it was asked. Filtering co-orbiting pairs is the analyst’s job, and the fastest filter is a minimum relative speed.

Collision probability is a statement about your covariance

Probability is only reported when you supply --ric-sigma-km:

orbitforge conjunction screen \
  --constellation iss.json --catalog stations.tle \
  --start 2026-06-25T19:29:47Z \
  --duration-hours 2 --threshold-km 10 \
  --hard-body-m 20 --ric-sigma-km 0.5,0.5,0.5

Holding the geometry fixed at a 1.164 km miss and varying only the assumed covariance:

Same encounter, same miss distance, and the probability spans 54 orders of magnitude.

Probability dilution

Read that table again and notice it is not monotonic. Probability peaks near a sigma of 0.55 km, roughly half the miss distance, and falls away on both sides.

• Small sigma. You know both orbits precisely, and you know they miss. Low probability, and genuinely reassuring.
• Large sigma. The probability mass is smeared across an enormous volume, so very little of it lands on the hard body. Low probability, and not reassuring at all.

Hard-body radius scales cleanly

Unlike covariance, this one behaves predictably. At a fixed 0.5 km sigma:

Probability scales with the square of the radius, as an area should. Doubling from 10 to 20 m multiplies it by 4; going from 20 to 50 m multiplies it by 6.25.

That is a check you can run yourself, and it is worth running once on any tool before trusting its probabilities.

A workable procedure

1. Screen at a generous threshold to see the population.
2. Discard events below a minimum relative speed, which removes co-orbiting pairs and self-matches.
3. Confirm your own satellites are not present in the catalog.
4. For survivors, obtain a real covariance rather than assuming one.
5. Compute probability across a range of plausible covariances and report the range, not a single value.
6. Treat a low probability from a large covariance as a tracking problem, not a safety result.

Checklist

1. Does the catalog contain your own satellites?
2. Did you filter on relative speed?
3. Is every probability quoted with the covariance behind it?
4. Is the covariance measured or assumed, and does the reader know which?
5. Is the hard-body radius realistic for both objects together?

Next steps

• conjunction screen reference for the full flag set.
• Import a real fleet for the epoch discipline TLE data requires.